Comment by 7777777phil

You're making a distinction the paper should address more directly. The classifier can't tell the difference between "this API design is fundamentally flawed because X" and "this company is terrible" (as noted in an earlier reply). Both register as negative by models trained on reviews and social media.

You're also right that HN's moderation probably removes hostile content quickly (which is why I prefer this platform to other roptions tbh). So the negativity we observe is mostly substantive critique rather than personal attacks.

That said, I'd push back a bit on whether this makes the finding less interesting. If anything, the opposite seems true. The fact that HN's "negativity" is constructive criticism, and that this criticism correlates with 27% higher engagement, tells us something about how technical communities value critical analysis over promotional framing. The classifier limitation is real (also see my other replies), but the engagement correlation holds whether we call it "negative sentiment" or "evaluative critique."

I'll add a limitations section to make the terminology clearer: "negative sentiment" as used here means evaluative criticism detected by SST-2-trained models, not personal attacks or toxic comments. Thanks for your feedback!